A domestic research team has developed AI technology that can screen out drugs likely to cause fatal side effects before clinical trials.
A team led by Professor Kim Sang-uk of the Department of Life Sciences and Graduate School of Convergence at Pohang University of Science and Technology POSTECH developed technology that uses AI to predict drug side effects that will appear in humans. The study was published online on the 28th (local time) in the international medical journal eBioMedicine.
In new drug development, there are many cases where drugs that pass preclinical tests such as cell or animal experiments show unexpected toxicity in humans. This is because biological responses differ between humans and animals. Chocolate is generally safe for humans, but can be poisonous to dogs. Conversely, a drug that is safe for mice is not necessarily safe for humans. One of the main reasons for new drug development failures so far has been this "inter-species difference."
Focusing on the genotypic-phenotypic difference (GPD), a biological difference among cells, mice, and humans, the researchers analyzed along three axes how target genes of drugs act differently in humans versus preclinical models. These axes are a gene's impact on survival (essentiality), tissue-specific gene expression patterns, and a gene's connectivity in biological networks.
When validated with data from 434 risky drugs and 790 approved drugs, GPD features were significantly associated with drugs that fail in humans due to toxicity. Predictive power improved substantially compared with looking only at chemical structure, with increases in both the metric for correctly finding toxic compounds and the metric for overall prediction accuracy. The developed AI model showed the best predictive performance compared with state-of-the-art models.
Furthermore, it proved practical in "chronological validation," which flags drugs likely to be withdrawn from the market due to toxicity. After training the AI using only drug information up to 1991, it predicted with 95% accuracy which drugs would be withdrawn from the market after 1991.
Kim said this was "the first attempt to numerically reflect the biological characteristics of preclinical models and humans," adding, "By combining AI with bioinformatics, the 'valley of failure' in new drug development can be greatly reduced, and the era of developing safe and effective new drugs for humans more quickly is not far off."
References
eBioMedicine (2025), DOI: https://doi.org/10.1016/j.ebiom.2025.105994